Inventory costs are a problem, but a bigger problem is production costs.
To create a PCB you need artwork. This includes the electronic files; the drill plots, the photoplots, the etch resist masks and the silkscreens. You then need the production tools for the PCB. You'll need a stencil for the screen-printing of the solder; a pick and place machine program (and you hope your PCB software can provide something that the P&P machine can read); if there's any through hole component you might want a cropping plate made.
A team goes through the prototyping process, using hand built small batch stuff. But then they need to convert this into production ready stuff. So the above costs happen when you hope -but are not sure- that you're ready to build a full scale batch. You get ten PCBs made for test, and discover a missing trace. That means you need either another set of stuff or to modify every board as it's produced.
Without very careful internal quality procedures and excellent engineering standards it'd be very easy for a company to burn through $10k before producing any saleable product.
It is difficult to follow lean startup principles with hardware products. The feedback loop to design, build and test hardware is long and expensive. Then once you've got a workable design you'll need a lot of capital to do a production run.
Before we can have a true renaissance of hardware startups there needs to a be a platform for cheap rapid prototyping with push button production scaling.
In effect, contract manufacturers can enable "push button production scaling." They'll source parts, do engineering, pack, and ship--basically whatever you pay them to do. A high tech copy machine. The interface could certainly use some work though...
There were plenty of web app startups long before Heroku and EC2 were available, these just lowered the bar even further.
Things like arduino can lower the friction between hardware and software design quite substantially for hackers.
I wouldn't necessarily expect startups to be rushing out to compete with the dishwasher any time soon but the point is that it now becomes much simpler to think about "hooks" in your software into the physical world.
The way you explain it means two things. First the investment needed to built the necessary production capacity, but that is not production cost (production cost are a result of this via hourly rates and costs per piece). The ability to cale from a prototype stage to serial production is crucial here, and yes damn expensive and risky. One wrong step and the company doesn't even have to care about inventories.
Inventory kicks in once the company has scaled successfuly. And then iventories are more curial to success than production costs in terms of company survival. Market entry is won with product costs low enough to make a profit. Once in the market you live and die by cash expanses, inventories are huge part of this. Espacially since hardware becomes outdated pretty fast, high levels of exess stock can basically mean death for a young company.
But all that is one of the schleps coming with hardware, and bulding such an operation from scratch is a really interesting thing to do.
I suspect this would be less of a worry if you focussed on low volume high value products. If you are selling solutions that are $10,000+ then you can afford a less efficient manufacturing process. I know a few people in my local area who are doing just that in some interesting niches.
To make a PCB you design artwork, usually from the circuit diagram. This artwork is transferred onto clear sheets. You overlay the clear sheet onto a board - a sheet of fibreglass, a sheet of copper, and a layer of UV sensitive etch-resist. You expose it to UV, you fix it, you wash off the etch-resist that isn't fixed (thus exposing copper) and then etch the exposed copper off, leaving the traces.
This process involves a couple of one-off costs (design of the PCB; transfer of that design to photo artwork). It also involves some setup costs that you get every time you need more product - setting jigs to the right size to hold the stuff. The artwork costs a couple hundred dollars, and that does not include any of the design and engineering costs, that's just the price if you hand a bunch of electronic data to a PCB house.
A stencil for screen-printing solder? That's steel, laser cut, can only be used for your PCB, and if you change the design of the PCB you might need a new stencil. Each stencil costs a couple hundred dollars.
A cropping plate is steel, or sometimes fibre glass, plate that has holes drilled where you have through-hole components fitted. An assembly-operative would have the cropping plate held in a jig, with PCBs on it, and they'd stuff the PCB. When they've fitted all the through hole components the plate is put on a machine which slides a blade under the plate and crops the leads. The operative then takes the PCBs off the plate, and sends them off for (probably) wave soldering. Each cropping plate is individual, and cannot be used for any other product. You'll need as many cropping plates as you want people to build PCBs. Each plate is going to be a couple of hundred dollars (steel) maybe a hundred dollar (fibreglass).
So far you haven't got any product that you can sell to the end customer! You've just got setup costs that you hopefully amortise over the number of product you produce.
You're right that an electronic assembly house would tend to outsource all this stuff to other people rather than doing everything in house.
Tooling waived when re-ordered emphasises what I've said - this is a one off up front cost that you pay before you get any product that can be sold to customers.
Note that I haven't said that PCB manf is hard. I've even said that companies don't do this stuff in house, but that they outsource it.
I'm pretty sure Sunstone doesn't charge for tooling if you use their proprietary software. You're locked in, but the costs aren't too bad for development. I'm sure getting the actual gerbers is a pain. ExpressPCB has a similar deal where they will make three index card size boards for $51 total. Not too bad, but they charge ~$60 for the gerbers.
If you're using "normal" software that can generate gerbers, there are other options. If you don't mind long turnaround and overseas fab, Seeed Studio will make boards for ~$1 each, for 5cmx5cm [1]. Slightly more for larger boards.
For hobby projects and prototypes, OSH Park can't be beat. Started as a board pooling coop for open hardware, they now send out batch jobs a few times a week to Amitron in IL and produce high quality boards (ENIG, with silk screen and soldermask) for $5/in^2 (for three copies, so a 1.5in^2 board would be $3.75 each)[2]. Their ordering process couldn't be better--upload gerbers, approve an on-screen rendering, specify the number of boards, and enter your billing info. No confusing options to go through. Seriously, try it. It's awesome.
So board fab is not that expensive anymore, and perhaps cheaper than when you last looked. The turnaround times aren't great for the inexpensive options, but with luck or good EDA it's possible to get by without too many iterations.
Board fab is over an order of magnitude cheaper than that these days. Yes, I have old circuit boards that I designed 15+ years ago that cost me $400 for two bare PC board prototypes, but today I can get the same two for about $50 these days, shipped express from somewhere in China.
To go even more extreme, I just ordered a handful of an existing design from OSH park just to see what their quality is. Three tiny boards cost me $5. Five bucks!!!
To create a PCB you need artwork. This includes the electronic files; the drill plots, the photoplots, the etch resist masks and the silkscreens. You then need the production tools for the PCB. You'll need a stencil for the screen-printing of the solder; a pick and place machine program (and you hope your PCB software can provide something that the P&P machine can read); if there's any through hole component you might want a cropping plate made.
A team goes through the prototyping process, using hand built small batch stuff. But then they need to convert this into production ready stuff. So the above costs happen when you hope -but are not sure- that you're ready to build a full scale batch. You get ten PCBs made for test, and discover a missing trace. That means you need either another set of stuff or to modify every board as it's produced.
Without very careful internal quality procedures and excellent engineering standards it'd be very easy for a company to burn through $10k before producing any saleable product.